Research on 2D/3D registration algorithm based on cross-attention mechanism and uncertainty pose regression loss
MENG Ruoyu
CHEN Chunxiao
XIAO Yueyue
WANG Kunpeng
Abstract：To address the problems that the existing 2D/3D image iterative registration method is time-consuming and easy to fall into local optimization,we proposed a registration method for 2D X-ray images and 3D computed tomograph(CT)images based on Swin Transformer and cross-attention.Firstly,the cross-attention mechanism was introduced into Swin-Transformer self-supervised posture regression model to explicitly capture the similarity information between input images,and achieve 2D/3D image registration.Then,a double geodesic loss based on homoscedastic uncertainty was proposed to compute pose distances,which realized an effective regression for a wide range of pose offsets.The 2D/3D image registration experiments was conducted on the public dataset DeepFluoro,the mean target registration error(mTRE)and registration success rate(RSR)were(5.31±4.29)mm,88.25％,respectively.The ex-perimental results show that this method can effectively improve the image registration accuracy and RSR.
Keywords：Robot-assisted surgerySurgical navigation2D/3D image registrationPose regressionCross-attention
Publication Date：2024-12-28
Online Publishing Date：2026-09-11(First online date of this platform, not the publication date of the document)
Pages：6( 456-461 )